MULTISENGE: Pixel-Based Time Series Processing & Analysis

This repository contains the pipeline to process, structure, and analyze satellite image time series (SITS) from the MULTISENGE dataset. It handles the extraction of pixel-level time series, organizes them into machine learning-ready cross-validation folds, calculates global statistics, and performs temporal correlation analyses.

Data Availability

The code relies on the MULTISENGE dataset. You must download the necessary raw data (Sentinel-2 images, Ground Reference, and Labels) from Zenodo:

Dataset: MULTISENGE (Zenodo)

To run this pipeline, ensure you have the following components extracted locally:

• Sentinel-2 Images (s2_images)
• Ground Reference (GeoTIFF masks)
• Labels (JSON files describing the patches)

🛠 Installation

To simplify execution, this project uses a pyproject.toml configuration.

1. Clone the repository:

   git clone [https://github.com/your-username/multisenge-processing.git](https://github.com/your-username/multisenge-processing.git)
   cd multisenge-processing

2. Install in editable mode: Run the following command to install dependencies and register the multisenge executable on your system:

   pip install -e .

⚙️ Configuration (Mandatory)

The project uses Hydra for configuration management. All settings are defined in config/hydra_config.yaml.

Variables you MUST update

Before running any command, open config/hydra_config.yaml and update the dataset.paths section with the actual locations of your downloaded data.

You must provide valid paths for these 4 variables:

1. folds_csv: Directory containing your split CSVs (fold_0.csv, etc.).
2. labels: Directory containing the JSON label files.
3. ground_reference: Directory containing the GeoTIFF masks.
4. s2_images: Directory containing the Sentinel-2 images.